Deep Learning Reveals Liver MRI Features Associated With PNPLA3 I148M in Steatotic Liver Disease

Abstract

Background: Steatotic liver disease (SLD) is the most common liver disease worldwide, affecting 30% of the global population. It is strongly associated with the interplay of genetic and lifestyle-related risk factors. The genetic variant accounting for the largest fraction of SLD heritability is PNPLA3 I148M, which is carried by 23% of the western population and increases the risk of SLD two to three-fold. However, identification of variant carriers is not part of routine clinical care and prevents patients from receiving personalised care.

Methods: We analysed MRI images and common genetic variants in PNPLA3, TM6SF2, MTARC1, HSD17B13 and GCKR from a cohort of 45 603 individuals from the UK Biobank. Proton density fat fraction (PDFF) maps were generated using a water-fat separation toolbox, applied to the magnitude and phase MRI data. The liver region was segmented using a U-Net model trained on 600 manually segmented ground truth images. The resulting liver masks and PDFF maps were subsequently used to calculate liver PDFF values. Individuals with (PDFF ≥ 5%) and without SLD (PDFF < 5%) were selected as the study cohort and used to train and test a Vision Transformer classification model with five-fold cross validation. We aimed to differentiate individuals who are homozygous for the PNPLA3 I148M variant from non-carriers, as evaluated by the area under the receiver operating characteristic curve (AUROC). To ensure a clear genetic contrast, all heterozygous individuals were excluded. To interpret our model, we generated attention maps that highlight the regions that are most predictive of the outcomes.

Results: Homozygosity for the PNPLA3 I148M variant demonstrated the best predictive performance among five variants with AUROC of 0.68 (95% CI: 0.64-0.73) in SLD patients and 0.57 (95% CI: 0.52-0.61) in non-SLD patients. The AUROCs for the other SNPs ranged from 0.54 to 0.57 in SLD patients and from 0.52 to 0.54 in non-SLD patients. The predictive performance was generally higher in SLD patients compared to non-SLD patients. Attention maps for PNPLA3 I148M carriers showed that fat deposition in regions adjacent to the hepatic vessels, near the liver hilum, plays an important role in predicting the presence of the I148M variant.

Conclusion: Our study marks novel progress in the non-invasive detection of homozygosity for PNPLA3 I148M through the application of deep learning models on MRI images. Our findings suggest that PNPLA3 I148M might affect the liver fat distribution and could be used to predict the presence of PNPLA3 variants in patients with fatty liver. The findings of this research have the potential to be integrated into standard clinical practice, particularly when combined with clinical and biochemical data from other modalities to increase accuracy, enabling easier identification of at-risk individuals and facilitating the development of tailored interventions for PNPLA3 I148M-associated liver disease.

Keywords: deep learning; medical imaging process; single nucleotide polymorphism; steatotic liver disease.

FIGURE 1

Deep learning‐based workflow for predicting steatosis‐associated SNPs using PDFF maps and liver segmentation on abdominal MRIs. (a) PDFF maps were obtained by utilising fat‐water imaging software. However, during the image reconstruction process, water‐fat swaps can occur, leading to incorrect mapping of water and fat signals. To address this, a ResNet [27] was trained to remove these artefacts and filter out unsuitable images. (b) To calculate the liver PDFF value, we trained a U‐Net to segment the liver region. We further refined the segmentations by sigma clipping to reduce the impact of blood vessels, cysts and other factors that could potentially affect the PDFF value. Finally, we calculated the median liver PDFF value. (c) Patients with median PDFF values greater than 5% constituted the study cohort. Homozygous carriers and non‐carriers of each SNP were selected for the corresponding study cohort. The final scores were obtained on the test set after using five‐fold stratified cross‐validation and ensemble testing method. For genetic variants with high test scores, subgroup analyses were performed based on obesity, sex and age. This figure was created in BioRender. Chen, Y. (2024) BioRender.com/w67o014. Reproduced by kind permission of UK Biobank.

FIGURE 2

Exclusion criteria of our study cohort.

FIGURE 3

Visualisation of prediction‐contributing areas of PNPLA3 I148M carriers on liver MRIs in SLD group. (a) We selected true positive samples from the predictions of PNPLA3 I148M model by inference on the test set. Each sample includes the liver segmentation based on the magnitude image, an attention map highlighting the regions contributing to the prediction, and the distribution of hepatic fat in these areas. (b) AUROC and confusion matrix of PNPLA3 I148M model using the ensemble testing method. (c) The heatmap of AUROCs, resulting from the subgroup analysis based on obesity (BMI > 30), sex and age. This plot shows one of six attempts using different random seeds for data splitting. For each subgroup, we presented the AUROC from the current attempt, along with the minimum and maximum values (min‐max) observed across all attempts. This figure was created in BioRender. Chen, Y. (2024) BioRender.com/g20c852. Reproduced by kind permission of UK Biobank.

FIGURE 4

Comparative analysis of image modalities for detecting PNPLA3 I148M homozygosity on liver MRIs. To investigate the specific patterns that distinguish PNPLA3 I148M homozygosity in the liver region, we extended our analysis beyond the original magnitude images to include the water‐signal‐only image, fat‐signal‐only image, PDFF map, and R2* map. (a) The model using water‐signal‐only image achieved an AUROC of 0.67 (0.63–0.72), exhibiting the ability to detect negative samples but showing less effectiveness in identifying positive samples. (b) The model, using a fat‐signal‐only image, demonstrated an AUROC of 0.67 (0.63–0.73). Compared to the water‐signal‐only image, this model showed improved performance in detecting positive samples, although its ability to identify negative samples slightly decreased. (c) The model, using the PDFF maps, reached an AUROC of 0.65 (0.61–0.70). This model further enhanced the detection of positive samples. However, it frequently predicted false positives, likely due to a bias towards predicting positive samples. (d) The model using R2* map, which reflects the rate of signal decay influenced by local magnetic field inhomogeneities and often related to iron content in the liver, achieved an AUROC of 0.57 (0.52–0.62). This model produced a high number of false positives and exhibited a relatively low AUROC, suggesting that the R2* map patterns are not valuable for this specific task. Overall, both the fat‐signal‐only and water‐signal‐only images contain patterns that can distinguish PNPLA3 I148M homozygous carriers from non‐carriers. The performance of these models is comparable to that of the magnitude images. This figure was created in BioRender. Chen, Y. (2024) BioRender.com/j91q961.